Super-adapter lifts Qwen3.8-27B agents from 55.5% to 63.5% mean across 6 pairs
yb2698 · x · 2026-11-20
yb2698 shares results for the Qwen3.8-27B super-adapter with thinking on (xhigh), base → merged adapter:
- SWE-Bench Pro (hard 51): Hermes Agent 45% → 61%, OpenCode 47% → 67%, pi 74% → 74%
- Terminal-Bench 2.1: +4 to +6 points under each harness
- Mean across 6 pairs: 55.5% → 63.5%
The merged adapter also cuts tool calls by 19–48% in thinking mode depending on the harness.
More from coding & agent
- PowerShell creator's Meetup: start AI agents in an empty directory, ditch old cost estimates — dfinke · 2026-11-21
- Time awareness and async input: engineering patterns for personal agents — nikunjhanda · 2026-11-21
- Agent memory needs business ontologies, not smarter LLMs, argues founder — _AustinCalvert_ · 2026-11-21
- MCP Events extension finalized in beta, ready for client and server adoption — dsp_ · 2026-11-21
- Using Notion as an AI context layer with an external Claude skills-db — ivanhzhao · 2026-11-21
- Gaussian splats in Godot: full tutorial plus using Fable agent to plan a pixel apartment layout — XRarchitect · 2026-11-21